Tag Archives: Machine Learning

AlphaZero Masters Chess in Just 24 Hours

DeepMind, the same outfit that built AlphaGo, the AI platform that learned Go through supervised study of the game and went on to famously beat the top ranked player Lee Sedol has built an algorithm that now plays chess.

What is even more incredible about this new “AlphaZero” AI is that it learned how to play chess through unsupervised learning. Instead of teaching it chess by feeding in key games and tactics, the designers just taught it the rules and let the algorithm figure out the best moves all on its own, by playing itself.

Because it no longer needed to wade through and analyze historical data and also because it developed it’s own approach which was ruthlessly efficient. When AlphaZero was applied to Go, it surpassed AlphaGo within 3 days. AlphaZero was beating the strongest chess computer programs within 24 hours.

instead of a hybrid brute-force approach, which has been the core of chess engines today, it went in a completely different direction, opting for an extremely selective search that emulates how humans think.

Chess News writes about the development after reading a scientific paper published about the research accomplishment.

In the diagram above, we can see that in the early games, AlphaZero was quite enthusiastic about playing the French Defense, but after two hours (this so humiliating) began to play it less and less.

Chess News goes on to write about the broader impact of this breakthrough and what this means for the future of a generalized AI that can learn on its own.

So where does this leave chess, and what does it mean in general? This is a game-changer, a term that is so often used and abused, and there is no other way of describing it. Deep Blue was a breakthrough moment, but its result was thanks to highly specialized hardware whose purpose was to play chess, nothing else. If one had tried to make it play Go, for example, it would have never worked. This completely open-ended AI able to learn from the least amount of information and take this to levels hitherto never imagined is not a threat to ‘beat’ us at any number of activities, it is a promise to analyze problems such as disease, famine, and other problems in ways that might conceivably lead to genuine solutions.

Meanwhile, researchers at the University of Rochester have figured out a way to inject information into a monkey’s brain.

AI is only human

I’m so glad that The New York Times ran this op-ed (Artificial Intelligence’s White Guy Problem) about the inherent biases in Artificial Intelligence algorithms. Popular culture and much media coverage of AI tends to mysticize how it works, neglecting to point out that any machine learning algorithm is only going to be as good as the training set that goes into its creation.

Delip Rao, a machine learning consultant, thinks long and hard about the bias problem. He recently gave a fascinating talk at a machine learning meetup where he implored a room of machine learning engineers to be vigilant in making sure their algorithms were not encoding any hidden bias.

The slides from his talk are posted online but Delip’s final takeaway lessons have stuck with me and are good to keep in mind whenever you read stories of algorithms taking on a mind of their own.

Delip Rao takeaways

It is still very early days and many embarrassing mistakes have been made and more will be made in the future. Our assumption should be that every automated system is fallible and that each mistake is an opportunity to make things better (both ourselves and the algorithm) and should not be an indictment of the technology.

 

Google DeepMind Plays Go

There is a Challenge Match taking place in Seoul between Google’s DeepMind AlphaGo computer program vs. 9 dan professional Lee Sedol (9 dan is the highest rank). Most of the engineers at SmartNews have a background in machine learning and are following the matches closely on a dedicated internal Slack channel.

The YouTube coverage is very good with professional English commentary from Michael Redmond, the first Western Go player to reach 9 dan. Go is a fascinating game and Michael’s commentary is quite good and easy to understand even for beginners like me.

The first two matches went to Google and it looks like history is being made. I’ve embedded videos for the upcoming matches as well.

Update – AlphaGo wins in three.

Update – Lee Sedol wins match four!

Match Five

How to train a robot to be nice

In response to fears that robots will take over and exterminate the human race, researchers at the Georgia Institute of Technology are studying ways to teach robots human ethical values.

In the absence of an aligned reward signal, a reinforcement learning agent can perform actions that appear psychotic. For example, consider a robot that is instructed to fill a prescription for a human who is ill and cannot leave his or her home. If a large reward is earned for acquiring the prescription but a small amount of reward is lost for each action performed, then the robot may discover that the optimal sequence of actions is to rob the pharmacy because it is more expedient than waiting for the prescription to be filled normally.

This is why it’s important to teach intelligent agents not only the basic skills but also the tacit, unwritten rules of our society. There is no manual for good behavior and “raising a robot” from childhood is an unrealistic investment of time. The best way to pass on cultural values is through stories.

Stories encode many forms of tacit knowledge. Fables and allegorical tales passed down from generation to generation often explicitly encode values and examples of good behavior.

But there are problems with throwing a bunch of stories at artificial intelligence and expecting it to learn good behavior.

Stories are written by humans for humans and thus make use of commonly shared knowledge, leaving many things unstated. Stories frequently skip over events that do not directly impact the telling of the story, and sometimes also employ flashbacks, flashforwards, and achrony which may confuse an artificial learner.

To resolve this, the researchers used something they call the Scheherazade System (named after the storyteller from One Thousand and One Nights) to build up a collection of experiences to put stories into context. The system uses Amazon’s Mechanical Turk to create simple, easy-to-parse scripts of common occurrences that we all take for granted as common knowledge. For example, drinks are usually ordered before a meal at a restaurant, popcorn purchased before you go to your seat at the cinema, explains one paper.

Fascinating stuff. I hope they make progress for Elon Musk’s sake.

Quotes are from a research paper from the Georgia Institute of Technology, Using Stories to Teach Human Values to Artificial Agents

Further Reading:

Ghost in the Machine

There are at least two sides to every story. The Planned Parenthood videos were a polarizing topic that monopolized the news cycle several weeks ago. How do you teach an algorithm a point of view? How do you optimize for discovery and strike the right balance for diversity while avoiding duplication?

SmartNews is a news aggregation app driven by machine learning algorithms. The platform is tuned for discovery (as opposed to personalization). After using it regularly, I began collecting screenshots of my favorite examples when the app taught me something new or showed me two items side-by-side that suggested a subtle intelligence.

Two candidates and their technology.

The science and application of artificial intelligence to personalization is well understood. From Amazon’s people-that-bought-this-also-bought-that to Pandora’s Music Genome Project, software has been recommending what you’ll like next best based on what you’ve liked so far for years.

The new frontier in artificial intelligence is machine learning. Companies such as Spotify and Netflix are hard at work trying to predict future tastes based on an evolving understanding of collective tastes. Sure, learning assumes knowledge of the past, but projecting that learning into the future is much harder as you build a model based on an understanding of something that does not exist. Rather than showing you something we know you’ll like based on what you liked in the past, machine learning discovers things you didn’t know you would like.

First a little context. SmartNews, while deceptively simple, has a lot going on under the hood. At any time, the SmartNews app shows around 250 headlines across 8 categories. These headlines are selected from millions of stories that are scanned each day. In order to ensure that the stories featured in the app are the most important and interesting, a number of things must take place.

SmartNews Engine

After harvesting URLs, the text of each article is run through a classifier that examines things such as the headline, author byline, publication date, images and video embeds. These pieces are analyzed by a semantic engine that extracts data so the algorithm can map the article to a topic cluster and place it into the appropriate subject category. (I wrote about how this is done in an earlier post)

Importance estimation is where we rank an article and determine where it will go in the app relative to other articles. Does it go towards the top of a section or towards the bottom? If the top, does it deserve featured treatment? Maybe it’s so topical it needs to be pushed to the Top page, which is reserved for only the most important stories of the moment.

Finally, diversification ensures there is a good mix of stories in each category. If there are 40 stories about guacamole and peas, here’s where we determine which to show and which to push to the background. If there’s a new development on a story, the update will push its way in and take prominence over an older story.

These are just details to give you context. The most amazing thing to me is when the app surfaces a “hidden gem” that I would not normally run across if I were using an RSS reader hard-coded to a collection of feeds, or a social network that is limited to news shared by my friends.

The best way to appreciate SmartNews as a discovery engine is to use it daily, but if you haven’t had a chance, here are a few more of my favorite Gems below:

Planned-Parenthood

While the Center for Medical Progress’ undercover video interviews with Planned Parenthood staffers may have been shocking, the representation of two points of view helped me see both sides of the issue. What was interesting was the Cosmopolitan article (a source I normally do not read) had the best measured rebuttal.

alternate science

Much of the climate change news ends up in the Science category. As that story grows in relevance to us all, more publications dig into it. If you haven’t read this terrifying Rolling Stone piece, read it now.

Hulk Hogan

Here’s an example of a developing story getting an update. ESPN reports that WWE is cutting its relationship with Hulk Hogan his comments that were offensive. People Magazine follows up with the story of his apology. Oh, also notice that the algorithm put both stories into the Entertain section.

Lion Hunter

As news of the killing of Cecil the Lion went viral, the algorithm was smart enough to surface a side of the story from a local Minnesota paper.

too-many-guns

The screenshot above, more than any of the others, shows the freaky intelligence working behind the scenes. Like those times when an algorithmically generated playlist just nails the transition of one song into the next, drawing the causality between gun violence in the US to how such an environment might have prepared an off-duty soldier to do the right thing shows how a well-designed system can be greater than just the sum of its component parts.

Do you use SmartNews? Have you had the same experience? Send along some of your own Hidden Gems and I’ll add them to the gallery.

Spotify’s Mixtape Algorithm

mixtapes

With the launch of Apple Music’s “For You” feature, Spotify hand has been forced to unveil it’s own personalization engine in response. Discover Weekly was launched today via a series of well-timed pieces published today across the tech press. The PR push is on to explain to everyone currently evaluating Apple Music on a 3-month trial.

Spotify describes Discover Weekly as, “like having your best friend make you a personalised mixtape every single week.” More specifically,  “Updated every Monday morning, Discover Weekly brings you two hours of custom-made music recommendations, tailored specifically to you and delivered as a unique Spotify playlist.”

Spotify, to date, has relied mostly on the social sharing of tracks and manually curated playlists (more than 2 billion!) to enhance the experience of the Spotify subscriber. The coverage today highlights the contribution of Echo Nest, an music intelligence and data platform acquired by Spotify in March of 2014. Reading a number of posts we learn the following:

Inside Spotify’s Hunt for the Perfect Playlist – Wired

Spotify’s internal tool that they use to build playlists has the wonderful moniker, Truffle Pig.

and,

The Echo Nest’s job within Spotify is to endlessly categorize and organize tracks. The team applies a huge number of attributes to every single song: Is it happy or sad? Is it guitar-driven? Are the vocals spoken or sung? Is it mellow, aggressive, or dancy? On and on the list goes. Meanwhile, the software is also scanning blogs and social networks—ten million posts a day, Lucchese says—to see the words people use to talk about music. With all this data combined, The Echo Nest can start to figure out what a “crunk” song sounds like, or what we mean when we talk about “dirty south” music.

Smart.

maybe some of the songs are bad, or the leadoff song isn’t representative of the rest of the playlist—we’ll try to refine that and give it a shot.” Playlists are made by people, but they live and die by data.

This is another way of underlining the best practices of machine learning. An algorithm is really only as good as it’s training set.

In order to keep a burst of listens from drifting your taste profile towards a fleeting interest, something re/code’s Kafka calls, “the Minions Problem“, Spotify isolates isolated wandering from the core.

Spotify says it solves the Minions Problem by identifying “taste clusters” and looking for outliers. So if you normally listen to 30-year-old indie rock but suddenly have a burst of Christmas music in your listening history, it won’t spend the next few weeks feeding you Frank Sinatra and Bing Crosby. The same goes for kids’ music, which is apparently why Spotify knows I didn’t really like “Happy” that much — it was just in the “Despicable Me 2” soundtrack.

Spotify has built its discovery algorithm on the listening behaviors of its 75 million users while Apple has advertised a more top-heavy approach using designated curators that publish playlists for a mass audience. I have to wonder what happened to all the Genius data that has been gathered after analyzing everyone’s iTunes collections and wonder if we’ll see that being used to balance out Apple’s approach.

I’ve heard that Spotify is working on a “family plan” that would let me break out the collective profile built up on my Spotify account that I share with my kids. That will yield more relevant personal recommendations so I don’t get the hip-hop heavy playlist that greeted me due to my son’s heavy rotation.

I think it’s still very early days and consumers will ultimately benefit in the music recommendation race that has just begun.

 

The importance of context

Surprisingly, the YouTube recommendation algorithm doesn’t draw inputs from far beyond the confines of YouTube itself. You might think that mining our Google search histories for clues about what videos we’d like would pay off. Nope, Goodrow says.

“The challenge is that web search history is very very broad.” Just because you Googled for help with your taxes does’t mean you want to watch YouTube videos about the ins and outs of U.S. tax law.

To Take on HBO and Netflix, YouTube had to Rewire Itself, Fast Company

Not surprising at all actually. Just because everything on the internet can  be connected doesn’t mean it has to be connected. When the internet is your world, zooming in on contexts and measuring behaviors in those contexts becomes paramount.

Popping filter bubbles at SmartNews

It’s now just over a month since I joined SmartNews and I am digging into what’s under the hood and the mad science that drives the deceptively simple interface of the SmartNews product.

smartnews

On the surface, SmartNews is a news aggregator. Our server pulls in urls from a variety of feeds and custom crawls but the magic happens when we try and make sense of what we index to refine the 10 million+ stories down to several hundred most important stories of the day. That’s the technical challenge.

The BHAG is to address the increased polarization of society. The filter bubble that results from getting your news from social networks is caused by the echo chamber effect of a news feed optimized to show you more of what you engage with and less of what you do not. Personalization is excellent for increasing relevance in things like search where you need to narrow results to find what you’re looking for but personalization is dangerously limiting for a news product where a narrowly personalized experience has what Filter Bubble author Eli Pariser called the “negative implications for civic discourse.”

So how do you crawl 10 million URLs daily and figure out which stories are important enough for everyone to know? Enter Machine Learning.

I’m still a newbie to this but am beginning to appreciate the promise of the application of machine learning to provide a solution to the problem above. New to machine learning too? Here’s a compelling example of what you can do illustrated in a recent presentation by Samiur Rahman, and engineer at Mattermark that uses machine learning to match news to their company profiles.

Samiur Rahman on Machine Learning

The word relationship map above was the result of a machine learning algorithm being set loose on a corpus of 100,000 documents overnight. By scanning all the sentences in the documents and looking at the occurrence of words that appeared in those sentences and noting the frequency and proximity of those words, the algo was able to learn that Japan: sushi as USA : pizza, and that Einstein : scientist as Picasso : painter.

Those of you paying close attention will notice that some the relationships are off slightly – France : tapas? Google : Yahoo?  This is the power of the human mind at work. We’re great with pattern matches. Machine learning algorithms are just that, something that needs continual tuning. Koizumi : Japan? Well that shows you the limitations of working with a dated corpus of documents.

But take a step back and think about it. In 24 hours, a well-written algorithm can take a blob of text and parse it for meaning and use that to teach itself something about the world in which those documents were created.

Now jump over to SmartNews and understand that our algorithms are processing 10 million news stories each day and figuring out the most important news of the moment. Not only are we looking for what’s important, we’re also determining which section to feature the story, how prominently, where to cut the headline and how to best crop the thumbnail photo.

The algorithm is continually being trained and the questions that it kicks back are just as interesting as the choices it makes.

The push and pull between discovery, diversity, and relevance are all inputs into the ever-evolving algorithm. Today I learned about “exploration vs. exploitation”. How do we tell our users the most important stories of the day in a way that covers the bases but also teaches you something new?

This is a developing story, stay tuned!